FaultXAI: A Controlled Fault-Injection Study of Explanation Drift in ECG Classifiers
ID：86
Submission ID：478 View Protection：ATTENDEE
Updated Time：2026-07-22 16:09:48 Hits：17
Online
Start Time：2026-07-31 15:25 （Asia/Kolkata）
Duration：15min
Session：[S4] Computer Vision and Pattern Recognition » [S4-5] Computer Vision and Pattern Recognition
Video
No Permission
Presentation File
Tips: The file permissions under this presentation are only for participants. You have not logged in yet and cannot view it temporarily.
Abstract
Explainable AI (XAI) methods are increasingly deployed to interpret deep learning models in clinical decision support systems. However, the stability of these explanations under realistic signal degradation remains poorly understood. We introduce FaultXAI, a systematic framework for quantifying explanation drift the divergence between explanations of clean and corrupted inputs under controlled fault injection. We evaluate four clinically relevant fault types (Gaussian noise, baseline wander, lead dropout, segment dropout) across two benchmark ECG datasets: PTB-XL (12-lead, 21,837 records; 2,163 test records) and MIT-BIH (2-lead, 10,452 beats). Using Integrated Gradients and multiple stability metrics, our 5-seed experiments reveal: (1) additive faults exhibit strong cross-dataset agreement (r > 0.98), suggesting model intrinsic behaviour; (2) structural faults show dataset dependent effects; and (3) explanation drift can occur substantially even when predictive performance remains high (ρ < 0.7 at accuracy > 0.85). Our framework establishes reproducible benchmarks for
Keywords
Explainable AI, ECG classification, explanation drift, fault injection, robustness, deep learning, clinical decision support
Speaker
Comment submit